adaptive-trainer 0.0.6

A Hugging Face Trainer extension for adaptive loss and ideas learning.

Adaptive Trainer

A Hugging Face Trainer extension designed for more efficient and effective fine-tuning of language models using adaptive loss mechanisms and curriculum learning principles.

Features

• Adaptive Loss Calculation: Focuses training on tokens where the model is less confident or incorrect, rather than all tokens equally.
• Ideas Learning: Allows model to learn concepts and behaviour from training data rather than actual memorisation much like student learning.
• Learning Style Differentiation: Also allows specific datasets to be learnt with attention and concepts with ideas.Allows for different training objectives (e.g., "ideas" vs. "attention") based on dataset characteristics.
• Integration with Hugging Face: Built on top of the transformers library and Trainer API.
• More Features like parameter control and hyperparameter optimisation of loss for user comming soon

Installation

You can install adaptive-trainer using pip:

pip install adaptive-trainer

*For local directly from repo, clone and checkout to the required branch and run:

pip install .

if you want to install with optional dependencies like flash-attention:

pip install flash-attn --no-build-isolation
pip install bitsandbytes

Quick Start

from adaptive_trainer import train_adaptively, AdaptiveTrainer # Assuming you export AdaptiveTrainer too

# Define system prompts
system_prompts = {
    'both': "Your role as an assistant is ...",
    'ideas': "Your role as an assistant is ...",
    'attention': "Your role as an assistant is ..."
}

data_processing_function = lambda sample, context_mode: (sample['user'], sample['assistant'])
# Yes, for now you can't train more than 1 assistant response on one sample, if you need a histroy of conversation to train the next model response
# a trick to do is respond first element `sample['user']` with the assistant response included, you will need to however add the header start and close tokens for separating user and assistant texts mannually.

# Configure datasets
datasets_config = {
    'ideas': ["user/my-ideas-dataset:|:K"],
    'attention': ["user/my-attention-dataset"],
    'both': [],
    'misc': [],
    'data_processing_function':data_processing_function,
    # 'data_processing_function_attention':None,
    # 'data_processing_function_ideas':None
    # 'data_processing_function_dataset_name':None
}
# The above will trim the `user/my-ideas-dataset` till top K values (after filtering by max_token_length) and similar thing can be done for attention datasets too just add :|:K at the end of dataset name and it will only use top K rows of the dataset..
# The validation set will be proportionately be scaled down = max(K*len(val)/len(train), 10) if len(val)>10 not scaled otherwise
# Dataset preprocessing function which is expected to convert dataset sample row to user and assistant response strings.. is taken in the following fallback order:
# data_processing_function_dataset_name -> data_processing_function_{ideas/attention} -> data_processing_function
# **IMPORTANT** : passing data_processing_function_dataset_name with value None will not use fallback functions but rather used internal hard coded preprocessing function for different dataset patterns case to case.
# The hardcoded preprocessing might fail and raise error, therefore please pass the appropriate data_processing_function

dataset_kwargs = {
    'user/my-ideas-dataset:|:K': {'data_dir'=None, 'context_mode':'<parameter_value_for_dataset_preprocessing_function>', 'train_split':'<train_split_name>', 'val_split':'<val_split_name>'},
    'user/my-ideas-dataset': {}, # Both keys are valid and should work
    'user/my-attention-dataset': {}
}
# If you want to added custom arguments to `load_dataset` function from `datasets` library
# optional can be `None`
# train_split, val_split -> if the dataset has splits available they can be passed here, dataset is expected to have at least one split at high level when loaded

dataset_specific_system_prompts = {
    'user/my-ideas-dataset:|:K': 'Dataset_specific_prompt',
    'user/my-ideas-dataset': 'Dataset_specific_prompt', # Both keys are valid and should work
    'user/my-attention-dataset': 'Dataset_specific_prompt'
}
# dataset specific system prompt will get appended to master system prompt: system_prompt+dataset_specific_prompt
# optional can be `None`

# Configure training parameters
# Below are default values for all currently supported user_inputs:
training_config = {
    'run_name': 'my_adaptive_run',
    'wandb_project': 'my-adaptive-experiments',
    'wandb_entity': None
    'max_length_token': 4096, # datasets will be filtered with samples (user+assistant text) less than this value
    'padding_side': 'left',
    'batch_size': 4,
    'learning_rate': 2e-5,
    'gradient_accumulation_steps': 16,
    'num_epochs': 3,
    'attn_impl': None, # will use default implementation `sdpa` if available or `eager` otherwise
    'save_total_limit': 5,
    'fp16': True,
    'gradient_checkpointing': True,
    'eval_strategy': 'steps',
    'save_strategy': 'steps',
    'local_rank': -1,
    'logging_steps':10,
    'weight_decay': 0.01,
    'warmup_steps':10,
    'eval_steps': 200,
    'save_steps': 300,
    'eval_batch_size': 4, # defaults to batch_size
    'optimizer': 'paged_adamw_8bit',
    'use_liger_kernel': False,
    'eval_on_start': False
}

# Configure adaptive loss parameters
adaptive_loss_config = {
    'top_k': 8 # Currently only top_k parameter is allowed for user_control in training
    'adaptive_log_steps': 100 # after every these many training steps some adaptive training metrics will be logged
}
# It is expected that a lower value of top_k will lead to stricter learning possibly leading previous learned behaviour
# being forgotten and a higher value of top_k will allow more lienient learning
# top_k=1 can cause ideas learning to behave the same as attention learning

# Run training
model_path = train_adaptively(
    model_name="model_to_train",
    datasets_config=datasets_config,
    datasets_kwargs =datasets_kwargs,
    dataset_specific_system_prompts=dataset_specific_system_prompts,
    output_dir="./my_finetuned_model_adaptive",
    huggingface_repo="your_username/my_finetuned_model_adaptive", # Optional, if provided and the environ variable HF_TOKEN is detected then model will be uploaded
    system_prompts=system_prompts,
    training_config=training_config,
    adaptive_loss_config=adaptive_loss_config
)

print(f"Training complete. Model saved to {model_path}")

Contributing

... (Details on how others can contribute) not yet revealed ... Open issues as of now

License

This project is licensed under Apache License Version 2.0